GigitAI at SemEval-2026 Task 8: Hybrid Sparse-Dense Retrieval and Zero-Shot Generation for Multi-Turn Conversational RAG

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Natural Language Processing · Depth: Expert, medium

Summary

GigitAI presented its system for SemEval-2026 Task 8 (MTRAGEval), focusing on multi-turn conversational RAG. Their approach integrates hybrid retrieval, combining SPLADE-v3 sparse representations with dense embeddings using Reciprocal Rank Fusion, a fine-tuned cross-encoder reranker, and zero-shot LLM generation via Claude Opus 4.5. The team systematically evaluated 56 retrieval configurations across 4 domains and 5 generation strategies with 5 different LLMs. Key findings indicate that SPLADE-v3 with dataset rewrites significantly surpasses BM25, and simple zero-shot prompting performs comparably to more complex strategies like Self-RAG and CRAG. Performance varied considerably based on query answerability. On the test set, GigitAI achieved rank 5/29 for end-to-end RAG (H=0.5564), rank 7/26 for generation (H=0.7495), and rank 13/38 for retrieval (nDCG@5=0.4782). The system excelled on answerable queries (H=0.685) but showed degradation on underspecified ones (H=0.254).

Key takeaway

For Machine Learning Engineers developing multi-turn conversational RAG systems, consider prioritizing robust hybrid retrieval over complex generation strategies. Your efforts on sparse retrieval, specifically with SPLADE-v3 and dataset rewrites, can yield significant performance gains over BM25. Furthermore, simple zero-shot prompting with models like Claude Opus 4.5 can be as effective as more sophisticated methods. Focus on improving handling of underspecified queries, as this significantly impacts overall system reliability and user experience.

Key insights

Hybrid sparse-dense retrieval and simple zero-shot LLM prompting achieve competitive multi-turn RAG performance, with query answerability being a critical factor.

Principles

• SPLADE-v3 with rewrites beats BM25.
• Simple zero-shot prompting rivals complex RAG.
• Query answerability impacts RAG performance.

Method

The system uses hybrid retrieval (SPLADE-v3 + dense embeddings via Reciprocal Rank Fusion), a fine-tuned cross-encoder reranker, and zero-shot LLM generation with Claude Opus 4.5 for multi-turn conversational RAG.

In practice

• Implement SPLADE-v3 for sparse retrieval.
• Opt for simple zero-shot prompting.
• Address underspecified queries proactively.

Topics

• Multi-turn RAG
• Hybrid Retrieval
• SPLADE-v3
• Zero-Shot LLM Generation
• Reciprocal Rank Fusion
• Conversational AI

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Beyond Semantic Search: The Complete Guide to Advanced RAG with Milvus | Sagar Patil — Advanced RAG requires multi-stage retrieval, combining hybrid search with techniques like query rewriting and reranking for production reliability.
• An Efficient Approach for Answering Not Readily Attainable Questions for RAG-based Applications — This paper proposes an efficient approach for RAG systems to answer questions not readily found in retrieved data.
• RCEM: Robust Conversational Search EMbedder in Distributional Shift — RCEM distills LLM query rewriting into an embedder for robust, low-latency conversational search without explicit rewriting.
• Uncertainty-Aware Hybrid Retrieval for Long-Document RAG — UMG-RAG adaptively fuses multi-granularity, multi-expert retrieval based on query-specific uncertainty to create compact, coherent RAG contexts.
• Evaluation of Chunking Strategies for Effective Text Embedding in Low-Resource Language on Agricultural Documents — Effective text chunking significantly improves RAG performance in low-resource, morphologically complex languages.
• Rethinking RAG in Long Videos: What to Retrieve and How to Use It? — VideoRAG performance improves by using chunk-adaptive retrieval and reranking, addressing limitations of existing benchmarks and single-configuration approaches.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.